Mac dictation privacy: what actually leaves your Mac

Apple Dictation has two processing paths, and which one you get depends on your Mac and your language rather than on anything you choose. Apple's own Mac User Guide page for Dictation tells you to check Keyboard settings to see whether you need an internet connection, or whether your voice inputs and transcripts are processed on your device and not sent to Siri servers. That sentence is the whole privacy answer, and it is buried in a settings pane.

Where to look: Apple menu › System Settings › Keyboard › Dictation. If the pane says your audio and transcripts stay on the device, you are on the local path. If it does not, you are on the server path, and your speech is being transcribed in an Apple data center.

The split is not arbitrary. Apple's macOS feature availability page lists general Dictation in a longer list of languages than the one it lists for on-device and modeless dictation, and it notes that the on-device version requires a download of speech models. So on-device dictation is a subset: real where it applies, absent where it does not. Which language you dictate in decides your privacy posture, and how uneven that gets by Mac generation is spelled out in whether Apple Dictation is on-device.

Apple's Siri, Dictation and Privacy page is short, and it is the only document that actually answers the question. Four points from it, in Apple's own terms:

• Server-processed dictation is not stored by default. The things you dictate are sent to and processed on the server, but will not be stored unless you opt in to Improve Siri and Dictation.
• Related request data is stored regardless. Apple stores the category of request, whether it completed successfully, device specifications, device configuration, performance statistics, and, if Location Services is on for Siri, the approximate location of your device at the time.
• Nothing is tied to your Apple Account. Request history is associated with a random, device-generated identifier that rotates multiple times per hour and is not tied to your Apple Account or email address.
• Two years is the retention figure. Apple may retain and use that data for up to two years to develop and improve Siri, Dictation, Search and limited other language processing functionality.

Read that page closely and one thing does need holding at arm's length. The paragraph about Dictation says what you dictate will not be stored without the opt-in. Two paragraphs later, the same section says request history, which includes transcripts, may be retained for up to two years under a rotating identifier. Both sentences are in the same section, which covers Siri and Dictation together, and Apple does not spell out how they reconcile for Dictation on its own. It is not a contradiction so much as one document doing two jobs, but if you were hoping for a single unambiguous line to hand to a compliance reviewer, it is not there.

One thing Apple is unambiguous about: a history of your Siri and Dictation interactions is stored on your device. That is why there is a delete button, which is further down this page.

Dictation on a Mac has never been only about your voice. Apple's current privacy page says that when you use Siri and Dictation, your device may send other contextual data to Apple, and the first item on that list is contact names, nicknames and relationships, for example "my dad", if you set them up in your contacts. The reason is ordinary engineering: a speech model that knows the names in your address book gets those names right. The reason it matters is that the names of people you know are a different category of data from the sentence you just dictated.

This is not new, and it used to be much louder. When Dictation arrived on the Mac in 2012, macOS put a notice in front of you before you could use it. As Macworld reported at the time, the notice read: "When you use the keyboard dictation feature on your computer, the things you dictate will be recorded and sent to Apple to convert what you say into text. Your computer will also send Apple other information, such as your first name and nickname; and the names, nicknames, and relationship with you (for example 'my dad') of your address book contacts."

Fourteen years later the disclosure is still there, and it has moved. It is a bullet in a legal page rather than a dialog you have to dismiss, and the words "your address book" have become "your contacts". Nothing has been hidden. It has just stopped being something you read.

The honest limit of what I can tell you: Apple's on-device sentence is scoped to audio and transcripts, and says those are processed on your device and not sent to Apple servers. The contextual-data sentence is scoped to when you use Siri and Dictation, with no on-device carve-out attached. Neither sentence tells you which one wins when on-device dictation is active, and Apple does not publish that detail. So if you want a definitive answer for the local path, the documentation does not give you one. That is worth knowing before you rely on it.

This is the setting that changes the answer most, and it is opt-in. Apple's Improve Siri and Dictation privacy page spells out what turning it on collects: audio of your interactions with Siri and Dictation, along with transcripts and related request data such as the category of request, device specifications, performance statistics and approximate location.

Then the part people react to. A subset of that data may be reviewed by graders, and Apple states that audio review is conducted only by Apple employees. Apple may retain and use the data for up to two years, and the small subset of interactions that has been reviewed may be kept beyond two years. As with everything else, it is associated with the random rotating identifier, not your Apple Account.

You can check what you agreed to. The switch lives under Apple menu › System Settings › Privacy & Security › Analytics & Improvements, where it appears as Improve Siri & Dictation. Most people were asked once, during setup, in a screen offering "Share Audio Recordings" or "Not Now", and do not remember which they tapped.

To clear what is already there, go to Apple menu › System Settings › Apple Intelligence & Siri, then Delete Siri & Dictation History. Two caveats from Apple's own instructions: you need an internet connection to do it, and deleting the history does not change the choice you made about sharing audio recordings. Those are two separate actions and doing one does not do the other.

For context on why people are wary: in January 2025 Apple agreed to pay $95 million to settle a proposed class action over Siri recordings captured by accidental activations, and published a statement on its Siri privacy commitment saying it has never used Siri data to build marketing profiles, never made it available for advertising, and never sold it to anyone for any purpose. Worth being precise about the scope: that case concerned Siri activations, not Dictation, and Apple did not concede the claim. It belongs in a risk assessment as context, not as evidence about dictation.

If you have installed a dictation app, that app is now the bigger question, not macOS. Apple Dictation at least has an on-device path and a public two-year ceiling. A cloud dictation app has neither by design, because the speech model runs in a data center and the audio has to get there for the product to work at all.

Wispr Flow and Aqua Voice are the two cloud examples worth naming. Both offer a privacy mode, and both describe it as something you switch on rather than the default. Wispr Flow's policy says that enabling Privacy Mode and disabling Private Cloud Sync gives zero data retention, with no dictation data used for training and no audio or transcripts stored on its servers, and that without Privacy Mode your data may be used to improve the product's features and AI models. Aqua Voice's policy says that for users with Privacy Mode disabled it may securely store transcript data on its servers to the extent necessary to improve the product.

I could not find a published retention period in either policy, so I am not going to give you one. What is true architecturally, and does not depend on any policy line, is this: when the audio leaves the machine, your privacy is a setting plus a contract plus the vendor's future decisions. All three can change without you noticing, and none of them is something you can verify from your own Mac. That is the difference between privacy that is promised and privacy that is structural, worked through in architectural versus contractual privacy.

There is a second leg people miss even with local apps. An app can run the speech model on your Mac and still send the finished transcript to a cloud LLM the moment you turn on an AI cleanup or summarize feature. If you have got past the risk assessment and now want to know which specific apps pass on both legs, that is a different job and it has its own page: which Mac dictation apps work offline and keep your audio private does the app-by-app filtering, while this page stays on what happens to your voice in the first place.

You do not have to take anyone's word for any of this, mine included. Four checks, in order, and none of them needs a tool you do not already have.

1. Read the indicator. Apple menu › System Settings › Keyboard › Dictation. Look for whether it says your audio and transcripts are processed on the device. That is Apple telling you which fork you are on for the language you have selected.
2. Cut the network and dictate. Macs have no Airplane Mode, so do it by hand: turn Wi-Fi off in Control Center, unplug Ethernet, and switch off any iPhone hotspot or tethered connection. Then put the cursor in a text field and dictate a sentence. If the words appear, the speech model is on your Mac. If it fails or hangs, it needed the server.
3. Watch the traffic. Open Activity Monitor, click the Network tab, sort by Data sent/sec, then dictate with the network back on. This is the check that also covers any third-party dictation app you have installed, because the app is right there in the list either sending or not.
4. Check what you opted into. Apple menu › System Settings › Privacy & Security › Analytics & Improvements, and look at Improve Siri & Dictation. Then clear the stored history under Apple Intelligence & Siri, as above.

Do that once and you will know more about your own dictation setup than any article can tell you, because you will have tested your Mac, your language and your apps rather than the general case.

A proportionate read, because the honest answer is not "a lot" for most people. Apple's posture on dictation is better than the industry norm: no audio retained by default, no link to your Apple Account, a rotating identifier, a published retention ceiling, an on-device path that genuinely exists, and a delete button that works. If you are dictating a shopping list or a Slack reply, this is not a problem you need to solve.

The cases where it does matter share one feature, which is that you cannot afford to answer "I would have to check":

• Privileged or confidential text. Client matters, clinical notes, HR cases, anything under a duty of confidence. See is Apple Dictation HIPAA compliant for how that plays out in a US healthcare review.
• A language off the on-device list. Your posture then is the server path, every time, and no setting changes it.
• An audit or information-governance questionnaire. "On-device for this Mac in this language" is a true answer that reviewers find hard to accept, because it is conditional.
• A shared or managed Mac where you did not make the setup choices and cannot be sure what was agreed to.

In those cases the fix is not a better setting. It is removing the fork, so there is no server path to reason about and nothing whose behavior can change in the next macOS release. The general shape of that argument is in what on-device speech recognition is, and the everyday-use side of the built-in feature is in voice typing on Mac.

I make Parakeety, a one-person Mac app, so I have an interest in this question and you should read me accordingly. What I have tried to do here is quote Apple rather than characterize it, mark clearly where Apple's documentation stops answering, and hand you the checks so you can verify your own Mac instead of trusting a vendor page. Where I could not confirm a figure, I have said so and left the gap rather than fill it. If you want the background on why I built the app this way, there is why I built Parakeety, and the full network story for my own app, including what its license check sends, is on the security page.

If the whole exercise above ends with you having to check a settings pane to know where your last sentence went, the way out is dictation with no fork in it. Parakeety runs Parakeet TDT v3 on the Apple Neural Engine and has no cloud transcription path at all, not as a fallback and not behind an AI cleanup toggle, so there is no server behavior to audit and nothing to switch off. Hold the section key, talk, release; your words paste at the cursor in whichever app you were typing into. It needs Apple Silicon and macOS 14 or later. There is a free 7-day trial with no card required. After that it is $25 once.

Try Parakeety free →